feature selection in the independent component subspace for face

Pattern Recognition Letters 25 (2004) 1377–1388

www.elsevier.com/locate/patrec

Feature selection in the independent component subspacefor face recognition

H.K. Ekenel *, B. Sankur

Department of Electrical and Electronic Engineering, Bogazici University, Bebek 34342, Istanbul, Turkey

Received 2 October 2003; received in revised form 19 March 2004

Available online 11 June 2004

Abstract

This paper addresses the feature selection problem for face recognition in the independent component subspace.

While there exists, at least, energy principle to guide the selection of the principle components, the independent

components (ICs) are devoid of any energy ranking, and must therefore selected based on their discriminatory power.

In addition the independent component features can be selected starting from a much larger pool, or from a combi-

nation pool of ICA and PCA features. Four feature selection schemes have been comparatively assessed, and feature

subsets are tested on a face database constructed from CMU PIE and FERET databases. The discriminatory features

from larger pools are observed to be concentrated around fiduciary spatial details of the nose, the eyes and the facial

contour. Overall, face recognition benefits from the feature selection of ICA or PCA components and from the

combination of ICA and PCA feature pools.

� 2004 Published by Elsevier B.V.

1. Introduction

Face recognition has become one of the most

active research areas of pattern recognition since

the early 1990s. The interest on face recognition

is mainly fueled by the identification require-

ments for access control and for surveillance tasksagainst terrorism. This interest is still increasing,

since face recognition is also seen as an impor-

tant part of next-generation smart environments

(Pentland and Choudhury, 2000).

* Corresponding author. Tel.: +90-212-358-15-00; fax: +90-

212-287-24-65.

E-mail address: [email protected] (H.K. Ekenel).

0167-8655/$ - see front matter � 2004 Published by Elsevier B.V.

doi:10.1016/j.patrec.2004.05.013

Among the plethora of face recognition meth-

ods, the paradigm based on face appearance data,

template-based algorithms and their concomitant

subspace versions, such as PCA and LDA meth-

ods are the most popular (Turk and Pentland,

1991; Belhumeur et al., 1997). Recently a blind

source separation technique, called independentcomponent analysis (ICA) has been adopted

for face recognition as an alternative subspace

method. Face recognition algorithms, however,

encounter several difficulties due to changes of the

face appearances caused by such factors as occlu-

sion, illumination, expression, pose, make-ups

and aging. In fact the subsequent intra-individual

variability of face images can be larger than theinter-individual variability (Gong et al., 2000).

mail to: [email protected]

1378 H.K. Ekenel, B. Sankur / Pattern Recognition Letters 25 (2004) 1377–1388

One expects to surmount these difficulties, or at

least to mitigate their effect by judicious choice of

features that are insensitive to the variations in the

facial appearance. The purpose of this study is to

explore subsets of ICA and PCA features for face

recognition that increase recognition performance,and that are purportedly robust against expression

variations and differences in illumination. The

feature selection criterion used in PCA-based face

recognition is the eigenvalue variance. On the

other hand the independent components (ICs) are

devoid of any energy/importance ranking, there-

fore relatively high-dimensional feature vectors are

used in face recognition using ICA, i.e., 200dimensions (Baek et al., 2001; Bartlett et al., 1998;

Draper et al., 2003) with respect to PCA. To

handle this problem, proportion of variance (PoV)

has so far been the only feature selection technique

used in the literature on ICA face recognition

(Bartlett et al., 1998; Deniz et al., 2001; Yuen and

Lai, 2002; Havran et al., 2002). Besides using the

feature selection methods based on individualproperties of the features, like proportion of vari-

ance and best individual feature, we also utilize

more sophisticated feature selection schemes that

take into account the combinatorial properties of

the features such as sequential forward selection

and sequential floating forward selection. We as-

sess comparatively these four different feature

selection schemes on the basis of the improvementthey bring in the recognition performance. Fur-

thermore, since the first face recognition architec-

ture of ICA provides local features, we tried to

discover which one of these local features has more

discriminatory power, and thus contributes most

to the person identification.

The paper is organized as follows. In Section 2

we review briefly PCA and ICA techniques. Thefeature selection methods are presented in Section

3. Experimental results are discussed in Section 4

and conclusions are drawn in Section 5.

2. Subspace analysis methods

Face data, as obtained from the raster scanningof the face images, constitutes a very high-dimen-

sional space. However the intrinsic dimensionality

of the face space is known to be much smaller

(Gong et al., 2000), this despite the variations in

expression, pose and lighting. In fact the faces are

believed to be clustered on some low-dimensional

manifolds. The subspace techniques aim to reduce

the inherent excessive dimensionality of scanneddata to make the face recognition algorithms via-

ble and to capture or approximate the underlying

face manifolds. The most widely used subspace

analysis tools are the principal component analysis

(PCA), independent component analysis (ICA),

linear discriminant analysis (LDA) and their non-

linear varieties via kernel tools (Turk and Pent-

land, 1991; Belhumeur et al., 1997; Kim et al.,2002; Yang, 2002; Moghaddam, 2002).

In the sequel, lowercase and uppercase letters

denote scalar values, bold uppercase letters denote

matrices, and bold lowercase letters denote row

vectors.

2.1. Principal component analysis (PCA)

PCA aims to determine a new orthogonal basis

vector set that best reconstructs the face images, in

other words with the smallest mean-square error

for any given subspace dimensionality. These

orthogonal basis vectors, also called eigenfaces,

are the eigenvectors of the covariance matrix of the

face images. The most parsimonious eigenvector

set, say of dimension M , for the face reconstruc-tion problem is chosen as the subset of the M most

energetic eigenvectors, that is the eigenvectors

corresponding to the first M rank ordered eigen-

values.

Consider the K � D-dimensional face data ma-

trix X , where each D-dimensional row corresponds

to the lexicographically ordered pixels of one of

the faces, and where there are K face images. ThePCA method tries to approximate this face space

using an M-dimensional feature vector, that is

using M eigenfaces, where typically M �minðD;KÞ. These M eigenvectors span a face

subspace, such that kXk2 � kXVk2 is minimum,

where V is the D�M-dimensional matrix that

contains orthogonal basis vectors of the face space

in its columns. Once the projection bases V areformed, when a test image xt arrives, it is projected

onto the face subspace to yield the feature vector,

H.K. Ekenel, B. Sankur / Pattern Recognition Letters 25 (2004) 1377–1388 1379

rt ¼ xtV . The classifier decides for the identity of

the individual, according to a similarity score be-

tween rt and the feature vectors of the individuals

in the database fr1; r2 . . . ; rKg.

2.2. Independent component analysis (ICA)

Independent component analysis is an unsu-

pervised learning method based on high order

statistics. Briefly, ICA is the separation of inde-

pendent sources from their observed linear mix-

tures (Hyvarinen and Oja, 2000). The system

model of ICA is given as

X ¼ AS ð1Þwhere A denotes the mixing matrix, S denotes the

source matrix containing statistically independent

source vectors in its rows and X denotes the data

matrix. In the ICA method, the only information

we possess is the observations, and neither themixing matrix nor the distribution of the sources is

known. Under the assumptions that the sources

are statistically independent and non-Gaussian (at

most one of them can have Gaussian distribution),

we find the unmixing matrix W by maximizing

some measure of independence. In other words, a

separation matrix, W , is estimated, which, under

ideal conditions, is the inverse of the mixingmatrix A.

Y ¼ WX and W ¼ A�1 and Y � S ð2ÞIn the context of face recognition, the use of

ICA features was first proposed in (Bartlett et al.,

1998) and several other studies followed in (Liu

and Wechsler, 1999; Baek et al., 2001; Kwak et al.,

2002; Deniz et al., 2001; Yuen and Lai, 2002; Ding

et al., 2001; Havran et al., 2002; Draper et al.,

2003). In (Bartlett et al., 1998) two different ap-

proaches are presented for face recognition. Inthe first approach (called ICA1 architecture), the

face images are assumed to be a linear mixture

of an unknown set of statistically independent

source images. The source images obtained in this

architecture are spatially local and sparse in nat-

ure. In the second approach (called ICA2 archi-

tecture), the representation (weighting) coefficients

are assumed to be statistically independent. Inthis second architecture, while mixing coefficient

vectors are independent, source images tend to

have global face appearance, as in the case of

PCA. In contrast to the PCA method, where fea-

ture subset selection is based on energy criterion,

the selection of an ICA basis subset is not imme-

diately obvious since the energies of the indepen-dent components cannot be determined. One

typically first trims the minðD;KÞ-dimensional

space to M using a PCA stage, and then proceeds

to extract the ICA components from this M-

dimensional space.

The architecture of ICA1 and that of PCA have

analogous structures, as they are both based on the

statistical properties of the basis images. In ICA1,one tries to find statistically independent basis

images, whereas in PCA one extracts uncorrelated

basis images. ICA2 is based on the statistical

properties of the representation coefficients, thus it

has a different structure with respect to ICA1 and

PCA. In this work, both for a fair comparison

with PCA-based method and to explore the local

face regions that are appropriate for identifica-tion, the first architecture of the ICA is used.

In summary, the face recognition algorithm

using ICA1 architecture is as follows:

(i) Prior to ICA, PCA is performed on a train-

ing set and the M eigenvectors, associated

with the largest eigenvalues are selected to

form the matrix V .(ii) The training face images are projected onto

the PCA-based face space and the K �Mmatrix of their representation coefficients,

R ¼ ½r1jr2; . . . ; jrK �T, are obtained:

R ¼ X � V ð3Þ

(iii) An ICA analysis is performed on VT, where

face eigenvectors form the rows of this matrix;

the unmixing matrix, W , reveals the ‘‘ICA

faces’’

S ¼ W � VT ð4Þ

(iv) By using PCA representation coefficients, R,and independent basis images, S, ICA repre-

sentation coefficients of the faces in the train-

ing set, fa1; a2; . . . ; aKg, are calculated asfollows:


A ¼ R �W�1 ð5Þ

since VT ¼ W�1 � S, X ¼ R � VT, and

X ¼ R �W�1 � S. When a test image xt ar-

rives, it is projected onto the face subspace

to extract its PCA representation rt ¼ xtV ,

which is then multiplied with the inverse of

the separation matrix to yield the ICA featurevector at ¼ rt �W�1. This vector is compared

vis-�a-vis the feature vectors fa1; a2; . . . ; aKg of

the individuals in the face database.

3. Feature selection techniques

3.1. Rationale for feature selection for PCA/ICA

faces

The main goal of this study is the automatic

selection of the best feature subset for classifica-

tion purposes given a high-dimensional ICA fea-

ture vector. Recall that the feature selection

criterion used in PCA-based face recognition is the

eigenvalue variance. While this criterion yields themost compact set for the reconstruction, it does

not necessarily follow that it is optimal from face

recognition point of view. The ICA methods do

not have such an opportunity of feature selection,

in that the number of ICA features is pre-deter-

mined in the PCA stage of data processing. It is

conjectured that some feature selection scheme

focused on ‘‘recognition’’ rather then on ‘‘recon-struction’’ could augment face classification per-

formance. With this goal in mind, we have not

initially reduced the subspace dimensionality via

PCA to the target feature size, M . Instead we have

reduced the initial dimensionality minðD;KÞ downto an intermediate size M 0, M < M 0 < minðD;KÞ.The judicious choice of M 0, on the one hand,

should make the implementation of the ICAalgorithm viable and avoid overlearning effects

(Bartlett et al., 1998; Liu and Wechsler, 1999). On

the other hand M 0 should allow sufficient freedom

or richness of choice for the feature selection

algorithms to be effective. For example, if the

target feature vector will be 30-dimensional, we do

not effect a minðD;KÞ-to-30 PCA algorithm, but

use, for example, a minðD;KÞ-to-200 PCA reduc-

tion. Following this, the feature selection algo-

rithm proceeds to find the most discriminating

M ¼ 30 ICA features from the intermediate set of

M 0 ¼ 200.

We have comparatively assessed the recogni-

tion performances resulting from different featureselection algorithms, namely, proportion of vari-

ance (PoV), best individual feature (BIF), sequen-

tial forward selection (SFS) and sequential floating

forward selection (SFFS) techniques (Feri et al.,

1994). PoV has so far been the only feature selec-

tion technique used in the literature on ICA face

recognition (Bartlett et al., 1998; Deniz et al., 2001;

Yuen and Lai, 2002; Havran et al., 2002). PoVwould be the best approach if the features’ con-

tributions to the face recognition performance

were independent from each other. BIF is also

based on the same assumption as PoV, the differ-

ence being that, while in PoV the variances of

features are used, in BIF the classification perfor-

mances of the individual features are taken into

consideration. SFS technique is a special case ofthe well known ‘‘plus l ) take away r’’ methods,

and SFFS is an enhanced version of SFS. These

last two techniques search for the best performing

feature set while taking into consideration the

correlation between features.

3.2. Proportion of variance (PoV)

In this technique, for all feature points the ratio

of between-class variance to within-class variance

is calculated and the feature points that have high

ratio are selected. In this context ‘‘class’’ denotes

the ensemble of images belonging to one individ-

ual. Thus the mth feature of the qth face image

ðq ¼ 1; . . . ;QÞ of the kth ðk ¼ 1; . . . ;KÞ individualwill be expressed as fmðk; qÞ. The class mean will beobtained as

�fmðkÞ ¼1

Q

XQq¼1

fmðk; qÞ ð6Þ

while the average of this feature over the database

will be denoted as

f m ¼ 1

KQ

XKk¼1

XQq¼1

fmðk; qÞ ð7Þ

Fig. 1. Sample face images: four images of each of two indi-

viduals with differing illumination effects, expression or acces-

sories.


The ratio of between-class variance to within-class

variance for the mth feature point in the feature

vector is calculated as

rm ¼ rbetween;m

rwithin;m

¼PK

k¼1

��fmðkÞ � f m

�2PK

k¼1

PQq¼1 ðfmðk; qÞ � �fmðkÞÞ2

ð8Þ

This measure reveals out whether a change in fa-

cial appearance corresponds to a change in the mthfeature point or not. Thus we rank the featuresaccording to their rm ratio where frðmÞ is the mthranking feature, and select the highest ranking

M features, ffrð1Þ; frð2Þ; . . . ; frðM�1Þ; frðMÞg.

3.3. Best individual feature (BIF)

In this technique, classification performance of

each feature point is calculated separately, that is,on individual basis, and the features giving rise to

highest correct recognition rate are selected. Let Pmbe the classification performance of the mth fea-

ture point, that is, Pm is the probability of correct

detection using solely feature fm. Then the feature

points are ordered according to their individual

classification performances Pm and the first Mones, having highest Pm values, are selected.

3.4. Sequential forward selection (SFS)

In the SFS method, features are selected suc-cessively by adding the locally best feature point,

the feature point that provides the highest incre-

mental discriminatory information, to the existing

feature subset. The SFS technique starts as the

BIF by identifying the first feature that has the

highest discrimination power. It proceeds, how-

ever, by adding sequentially to the selected subset,

those features that contribute most to the classifi-cation performance on top of the already selected

ones. Thus, from a single initial BIF feature, the

SFS subset grows to a pair, to a triple. . .till anM-fold subset is found.

3.5. Sequential floating forward selection (SFFS)

SFFS is similar to the SFS in adding features tothe subset; however, in addition, it goes through

cleansing periods, in that features are removed

systematically so long as the performance im-

proves after pruning. Note that the performance of

the pruned subset is compared with that of a

previous one with the same population. This

stratagem helps to avoid the nesting effect, whichresults when one is stuck with a suboptimal subset.

4. Results and discussion

4.1. Experimental setup

We have used a face database constructed withimages chosen from CMU PIE and FERET da-

tabases (Phillips et al., 2000; Sim et al., 2002). In

our experiment set, there are 214 individuals each

having four different frontal face images, making a

total of 856 images. 584 out of the 856 images are

chosen randomly from the FERET database, fafb

image set. These FERET-based images mainly

differ in expression. The remaining 272 imagesderive from the CMU PIE database, which con-

tain variations in illumination. All the face images

are aligned with respect to the manually detected

eye coordinates, scaled to 60 · 50 pixels resolution

and histogram equalized. For each individual in

the set, two of their images that contain normal

facial expression and have frontal illumination are

used for training, and the remaining two imagesthat contain alternative facial expressions and

illumination from left and right sides are used for

testing purposes (Fig. 1).


Feature vectors are extracted via independent

component analysis and principal component

analysis schemes. The FastICA algorithm (Hyvari-

nen and Oja, 2000) is used to perform independent

component analysis. In either case we extracted 200-

dimensional feature vectors, that is, the original3000-dimensional image feature vectors consisting

of raw pixel values were reduced to, respectively,

200 PCA features conserving 96.46% of the energy

and 200 ICA features. Any further feature selection

was carried on these intermediate sets of 200.

We used the nearest-neighborhood classifier in

our feature selection study. We opted for the

nearest-neighbor method, as it is a powerful non-parametric classifier without any costly training

stage since otherwise one needs to re-design the

classifier repetitively for each selected feature

subsets. Finally we evaluated comparatively three

different distance metrics, namely, the L1 norm,

the L2 norm, and the normalized correlation

coefficient, defined as follows:

L1 : d ¼XMj¼1

jftraining;j � ftest;jj ð9Þ

L2 : d ¼XMj¼1

jftraining;j

� ftest;jj2

!1=2

ð10Þ

Fig. 2. Sample basis images: first row––eigenfaces; second row––typic

row––typical instances of 30 ICA faces obtained from 200 eigenfaces

CC : d ¼f training � f test

kf trainingk � kf testkð11Þ

In Fig. 2, sample ICA and PCA basis images are

shown. In the first row, samples of eigenfaces are

seen. The basis images in the second row are in-stances of ICA faces obtained by applying source

separation to 30 eigenfaces, whereas in the third

row we display basis images that are obtained by

separating 200 eigenfaces. It can be observed that,

as the number of independent sources increases,

the ICA images become more spatially localized

and sparse. On the other hand, when a small

number of sources are used, the ICA basis imagesresemble the eigenfaces.

4.2. Experimental results

In the sequel we present the classification results

as well as the cumulative matching scores (CMS)

of face images under various feature selection

schemes and distance metrics.

In Table 1, the correct classification perfor-

mances of ICA and PCA are given when, respec-

tively, 30 and 200-dimensional feature vectors are

used. As it has also been reported in the literature(Moghaddam, 2002), there are no significant per-

formance differences between ICA and PCA. As

al instances of 30 ICA faces obtained from 30 eigenfaces; third

.

Table 1

Correct classification performance of ICA and PCA

L1 L2 CC

ICA-30 75.47 74.30 73.60

PCA-30 74.77 74.30 73.60

ICA-200 80.61 79.67 78.50

PCA-200 81.07 79.67 78.50


one should expect, the correct recognition rategradually increases with the increase in the

dimension of the feature vector. For both feature

sets, the L1 norm gives the best results; hence the

L1 distance metric is taken into consideration

while performing feature selection. Recall that in

this experiment the most energetic 30 (200) PCA

components were selected as feature sets, and that

the ICA features were simply obtained by furtherprocessing these PCA components via the Fast-

ICA algorithm. The rationale for this choice is

explicated in Fig. 3 on the basis of the relative

Fig. 3. Energy percentage accounted by the eigenfaces.

Table 2

Performance results of the feature selection techniques, when the initi

combined case

Feature selection method PCA-30/200 I

PoV 74.30 7

BIF 74.07 7

SFS 77.34 8

SFFS 78.74 8

energy accounted for by the chosen number of

eigenfaces.

To prove the conjecture that a judiciously se-

lected subset from the individual feature sets or

from the merged set of both ICA and PCA faces

can yield better performance, we carried the fol-lowing experiments. In the first experiment, we

selected subsets of 30 features out of the initial

set of 200, individually for the PCA and ICA

cases using various selection algorithms. In the

second experiment, we merged the two sets to

constitute a superset of 400 dimensions (200

PCA+200 ICA) and then proceeded to select the

subset of 30 features from this augmented featureset. Note that before merging of the two feature

sets, they were rendered commensurate by vari-

ance normalization of the ICA and PCA feature

vectors. In Table 2, the performance values of

the selected feature subsets are shown. Sequential

selection methods perform better than the PoV

and BIF methods. The reason can be attributed to

the fact that, in classification, the class-condi-tional features are not independent. Therefore

combinatorial feature selection proves to be more

effective than selecting them based only upon

their individual performances. In other words,

selecting the feature points that have high indi-

vidual classification performances does not imply

the best subset for classification. For example, in

Fig. 4, the ratio of between-class variance towithin-class variance is shown. Although ICA

coefficients have higher ratio as compared to those

of PCA coefficients, this does not necessarily imply

higher classification performance when considered

jointly as can be observed from Table 2. These

results show that the correct combination of the

feature points in the selected subset is more

important than any selection based on individualproperties.

al feature set is reduced from 200 to 30. From 400 to 30 in the

CA-30/200 Combined 30/400

1.73 72.90

5.93 71.73

0.84 79.67

3.64 86.68

Fig. 4. The ratio of between-class to within-class variances.

Fig. 5. Cumulative matching score curves of SFFS.

Fig. 6. Selected PCA feature points by SFFS.

Fig. 7. Selected ICA feature points by SFFS.


One can observe from Table 2 that, in the

sequential selection techniques, ICA achieves

higher correct recognition rates as compared to

PCA. While trying to bring down the dimension-

ality of the PCA feature set proves detrimental, the

converse is true for the ICA set. As the ICA fea-

ture set is reduced from the original 200 compo-

nents down to dimension 30, the performanceimproves by 3%, that is from the score of 80.61 for

ICA-200 to the score of 83.64 for ICA-30 using the

SFFS method. This increase in classification per-

formance is a consequence of the removal of the

ICA basis images localized on those face regions

experiencing major changes in the appearance and

having less discriminatory information.

It is interesting to observe that the subset se-lected from merged 400 dimensional ICA and

PCA feature vectors perform even better than

separate ICA or PCA feature vectors, despite the

fact the ICA features were obtained from PCA

features via a linear method. Finally we can ob-

serve that the SFFS feature selection method

proves to be uniformly superior in all experiments.

In fact the ranking in increasing performance ofthe feature selection methods in this experiment is

as follows: PoV<BIF<SFS<SFFS. In Fig. 5,

cumulative matching score curves of SFFS method

are given. The horizontal lines in the figure rep-

resent the performance values of the 30 and 200-

dimensional PCA and ICA feature vectors

obtained without any selection. As can be seen

from the figure, in ICA, nearly a 20-dimensional

selected feature vector can outperform the initial

200-dimensional feature vector, or similarly, a 14-

dimensional selection of features is on a par with

the 30-dimensional feature performance.In Figs. 6–8, we show the indices of the selected

30 features from the original set of 200 PCA and

ICA features and from the combined set using

SFFS method. In Fig. 6, the features are ordered

according to the energy index, that is the size of

Fig. 8. Selected feature points from the combined set by SFFS:

note that the first 200 indices refer to the ICA features, and the

second set of features, with indices from 201 to 400 refer to the

PCA features.


their eigenvalue. In Fig. 7, the ordering is what

results from the ICA algorithm, after the PCAreduction stage. Finally note that in Fig. 8, the

ICA and PCA features are juxtaposed in a single

graph, such that the first 200 indices correspond to

the ICA features and the last 200 indices corre-

spond to the PCA features. Contrary to the com-

mon practice of selecting the most energetic PCA

features, as can be observed in Fig. 6, only two

Fig. 9. Selected ICA faces with SFFS method from

thirds of the PCA feature are from the first 100

points. This shows that some of the less energetic

PCA coefficients, which are normally discarded for

image representation and compression, may still

contain valuable discriminatory information. The

selected ICA features are spread more evenly overthe index range. Finally, it is interesting to observe

that the number of selected feature indices from

PCA and ICA feature pools are approximately

equal.

In Fig. 9, the selected ICA features are pre-

sented. One notices that these features are char-

acterized by sparse regions localized around the

eyes, on the face contours and at the nose. One caninterpret these regions in the selected ICA com-

ponents as being less sensitive to the variations in

the facial appearance and at the same time to

contain more discriminatory information. These

findings confirm the results of a recent study on

optimal Gabor kernel location selection (Gokberk

et al., 2003). Another observation from Fig. 9 is

that, these basis images resemble the receptor fields

a training set of normal face appearances.

Fig. 10. Samples of misclassified face images. First image in the

pair is the test image; second image in each pair is the matched

training image.


obtained in local feature analysis (LFA) (Penev

and Atick, 1996).

4.3. Discussion of the results

The rationale to resort to a feature selectionscheme was to improve the face recognition per-

formance and to mitigate the effects of intra-vari-

ation. We investigated the source of the attained

improvement (compare for example, ICA-30 per-

formances in Tables 1 and 2). Some of the face

image, which were erroneously classified without

feature selection in ICA, but were then correctly

labeled after feature selection are shown in Fig. 10,in pairs side by side the test image and erroneously

matched training image. As can be seen from these

face pairs, their appearances are very similar and

the individuals can be discriminated by only a

careful consideration of the local face regions.

Appropriate selection of the ICA basis images

provides the utilization of this valuable local dis-

criminatory information.

Fig. 11. Selected ICA faces with SFFS method from a training

A desirable aspect of the proposed scheme

should be the robustness from one face database to

another. In other words, one would like to

encounter the same or very similar subset of fea-

set of faces with expression and illumination variations.


tures, as one switches from one face database to

another one. To validate the usefulness of the

features we have selected, we perform ICA on a

different training set. We swap the images in the

test set with those in the training set. Thus, in this

new arrangement, the face images in the trainingset contain differences in expression and variations

in illumination, and the face images in the test set

contain normal facial expression and have frontal

illumination. The independent components that

correspond to the selected ICA feature points with

SFFS method are presented in Fig. 11. As can be

observed from Fig. 11, again the ICA basis images

that localize on the eye region, outlines of faceand nose are selected.

5. Conclusions

In this study we have explored feature selection

techniques on ICA and PCA bases for face rec-

ognition. Feature selection techniques are war-

ranted especially for ICA features since these are

devoid of any importance ranking based on energy

content as the PCA components. The study was

carried out on a face database that contains bothfacial expression and illumination variations. Four

different feature selection techniques were used

comparatively and the sequential floating forward

selection method was observed to be uniformly

superior in all cases, in that the maximum correct

classification rate was obtained with its feature

subset.

The major conclusion from this study was thatthe feature selection applied on ICA features def-

initely improves the recognition performance by

8.17%. Furthermore, if the features are selected

from the augmented pool of both ICA and PCA

features, the performance improvement becomes

11.21%. In other words, it pays to select ICA/PCA

features subset from a larger set of them, rather

than deciding a priori for the dimensionality of thefinal feature subset. Indeed, instead of selecting

features on the basis of the first M most energetic

PCA components or their ICA versions, it was

more beneficial to search for the subset resulting in

the best possible classification performance from a

larger pool. It is interesting to note that even the

PCA features benefited from this approach, when

30 of them were selected from the set of 200 via the

SFFS algorithm instead simply selecting the first

most energetic ones, as is commonly done in the

literature.

When the ICA features are selected from alarger initial set, we observe that the resulting

features are more localized, as the sharp dark or

bright spots in Figs. 9 and 11 indicate. In fact, not

surprisingly the majority of these local accents are

around the eyes and nose, as well as close to the

facial contours.

References

Baek, K., Draper, B.A., Beveridge, J.R., She, K., 2001. PCA vs.

ICA: A comparison on the FERET data set. In: Internat.

Conf. on Computer Vision, Pattern Recognition and Image

Processing, North Carolina.

Bartlett, M.S., Lades, H.M., Sejnowski, T.J., 1998. Independent

component representations for face recognition. In: Conf.

on Human Vision and Electronic Imaging III, San Jose,

California.

Belhumeur, P.N., Hespanha, J.P., Kriegman, D.J., 1997.

Eigenfaces vs. fisherfaces: Recognition using class specific

linear projection. IEEE Trans. Pattern Anal. Machine

Intell. 19 (7), 711–720.

Deniz, O., Castrillon, M., Hernandez, M., 2001. Face recogni-

tion using independent component analysis and support

vector machines. In: Third Internat. Conf. on Audio and

Video Based Biometric Person Authentication, Sweden.

Ding, P., Kang, X., Zhang, L., 2001. Personal recognition using

ICA. In: Eighth Internat. Conf. on Neural Information

Processing, China.

Draper, B.A., Baek, K., Bartlett, M.S., Beveridge, J.R., 2003.

Recognizing faces with PCA and ICA. Computer Vision

and Image Understanding 91 (1–2), 115–137.

Feri, F.J., Pudil, P., Hatef, M., Kittler, J., 1994. Comparative

study of techniques for large-scale feature selection. In:

Pattern Recognition in Practice IV. Elsevier, pp. 403–

413.

Gokberk, B., Irfanoglu, M.O., Akarun, L., Alpaydın, E., 2003.

Optimal Gabor kernel location selection for face recogni-

tion. In: Internat. Conf. on Image Processing, Spain.

Gong, S., McKenna, S.J., Psarrou, A., 2000. Dynamic vision:

From images to face recognition. Imperial College Press.

Havran, C., Hupet, L., Czyz, J., Lee, J., Vandendorpe, L.,

Verleysen, M., 2002. Independent component analysis for

face authentication. In: KES’2002 Proc., Italy, pp. 1207–

1211.

Hyvarinen, A., Oja, E., 2000. Independent component analysis:

Algorithms and applications. Neural Networks 13 (4–5),

411–430.


Kim, K.I., Jung, K., Kim, H.J., 2002. Face recognition using

kernel principal component analysis. IEEE Signal Process.

Lett. 9 (2), 40–42.

Kwak, N., Choi, C., Ahuja, N., 2002. Face recognition using

feature extraction based on independent component anal-

ysis. In: Internat. Conf. on Image Processing, Rochester.

Liu, C., Wechsler, H., 1999. Comparative assessment of

independent component analysis (ICA) for face recognition.

In: Second Internat. Conf. on Audio and Video Based

Biometric Person Authentication, Washington, DC.

Moghaddam, B., 2002. Principal manifolds and probabilistic

subspaces for visual recognition. IEEE Trans. Pattern Anal.

Machine Intell. 24 (6), 780–788.

Penev, P., Atick, J., 1996. Local feature analysis: A general

statistical theory for object representation. Network: Com-

putation in Neural Systems 7 (3), 477–500.

Pentland, A., Choudhury, T., 2000. Face recognition for smart

environments. IEEE Computer 33 (2), 50–55.

Phillips, P.J., Moon, H., Rizvi, S.A., Rauss, P.J., 2000. The

FERET evaluation methodology for face-recognition algo-

rithms. IEEE Trans. Pattern Anal. Machine Intell. 22 (10),

1090–1104.

Sim, T., Baker, S., Bsat, M., 2002. The CMU pose, illumina-

tion, and expression (PIE) database. In: Proc. IEEE

Internat. Conf. on Automatic Face and Gesture Recogni-

tion, Washington, DC.

Turk, M., Pentland, A., 1991. Eigenfaces for recognition.

J. Cognitive Sci. 3 (1), 71–86.

Yang, M.H., 2002. Kernel eigenfaces vs. kernel fisherfaces: Face

recognition using kernel methods. In: Fifth IEEE Internat.

Conf. on Automatic Face and Gesture Recognition, Wash-

ington, DC.

Yuen, P.C., Lai, J.H., 2002. Face representation using inde-

pendent component analysis. Pattern Recognition 35 (6),

1247–1257.

feature selection in the independent component subspace for face

Documents